A cautionary tale on the use of software

An occasional commenter to my blog wrote an entry on "Where is engineering going". One of the subtopics in that was the on use and abuse of computers. In particular the use of computers as a black box. I have a cautionary tale about using computers as a black box.

About ten years a nuclear facility, "the customer," was buying a deployment system from a vendor. The vendor was a well known name in nuclear operations. The deployment system was to insert a high pressure cleaning lance into a waste tank to break up sludge on the bottom the of the tank. Once the sludge was broken up it was to be removed.

The vendor had had a finite element analysis of the entire deployment system performed by an outside consultant. The consultant's analysis showed the deployment system design to be satisfactory.

Because of the seriousness of a failure the customer wanted an independent assessment by an outside third party of the deployment system provided by the vendor. The customer approached my employer to have the assessment performed and the job landed on my desk. I was not to redo the entire analysis but only to analyze the wrist segment of the deployment system which was deemed the most critical. The significance of the wrist was that its failure would result in either the deployment system being unretrievable from the waste tank or dropping the payload which could result in breaching the tank.

I recieved a 3-D CAD file of the wrist, some drawings of the overall mechanism, and a table of loads. The vendor, viewing this whole excersize as a waste of time, was not interested in cooperating further. The cooperation I recieved was only due to arm twisting on the part of the customer.

The mechanism was a complex 3-d mechanism, as opposed two a 2-d planar, mechanism. It took me a few days to decipher the drawings, figure out the static equilibrium equations for the wrist, and set up the finite element analysis model. I ran the model and .... portions wrist near the mechanism were shown to be at yield under static loading conditions with no factor of safety. I was quite dismayed. I automatically assumed I made a mistake.

I went back through my analysis, both the hand calculation to determine the loading on the part, and the finite element analysis - same result. I refined my model of the pin joints to more accurately calculate the stresses at the connection points. It made matters worse - the stresses were now shown to be slightly higher than yield.

I was quite aware of the consequences to the customer/vendor relationship once I reported my results. I had two engineers I worked with review what I had done and they agreed with my conclusions. At this point I was still convinced I had overlooked something. It simply did not seem that an experienced vendor could make this sort of design error. With a heavy heart I called the customer and reported my findings. The customer, of course, was also convinced I had made a mistake but requested that I fedex my calculations and analysis results. The next day the vendor called and after expressing some doubts about my abilities and my employer's reputation also requested a copy of my work.

I heard nothing more for almost a week. By this time I was convinced I was in the dog house when the customer called. They agreed with my results and, reluctantly, so did the vendor.

What had happened?

The vendor's consultant had inadvertently modelled the spherical and pin joints as transferring moment loads where in actuality they can only transfer forces. That error dramatically changed the stresser on the part. The factor of safety showed in the vendor's original analysis was non-existent.

What were the end results of this? The project lead engineer was fired. Although he did not do the analysis, he signed off approving it. The consultant was fired. The customer eventually backed out of the contract. Although this flaw was fixed. This and other technical difficulties led to the customer losing confidence in the vendor.

Ever since then I have been leery of automatically trusting results from computer analyses. This poses a real problem in today's ever shortening design cycles. The assumption is if the computer did it, it's correct. There is no patience for checking the results or for performing alternative manual calculations.

Technorati tags
mechanical design
FEA

How do I make this?

I am designing a material storage container. One of the design requirements is that it can latch closed to prevent inadvertent opening. I came up with a design that utilizes the locking pawl shown in the picture above. There is a locking pawl at one end, an actuation arm on the other end, and a shaft connencting the two ends. The question now arises, "How do I have such an odd looking part made?" If it can't be made the design isn't worth pursuing.

This is not a mass production part. I will only need about twenty of them. This eliminates processes such as casting, forging, or powder metallurgy.

I concluded that it could be made on a CNC milling machine. It can be machined out of a solid block of material. The shaft of the part will be parallel to the milling machine spindle. The trick is to locate a milling cutter that will allow cutting on its top, side, and bottom surfaces. I show how envision it being cut in the second picture. The stock from which the part is being cut is in red. The cutter is shown in blue. The clamping fixture and the spindle are not shown.



A more elegant way of making this part would be to use a lathe with a C-axis and live tooling, or perhaps a swiss-screw type machine. I am, however, constrained by what is commonly available in my area. CNC milling machines are common but only one shop locally has a lathe with a C-axis and live tooling.


Postscript: My original design which used the locking pawl shown above was discarded in favor of a much simpler mechanism that uses a bent sheet metal tab.

technorati tags
mechanical design

machining

fabrication

Using statistics to design a holder

I recently designed a bracket to hold some scintillation detectors. The detectors were not all uniformly sized. To determine the opening size to make for the detectors I followed the procedure outlined below.

I measured the detector widths from a sample and collected the data listed in the table 1. The data has a mean of 1.192 inches (3.0277 cm) and a standard deviation of 0.008071 inches (0.0205 cm).

Table 1. Detector sizes.
data (in) data (cm)
1.18 2.9972
1.184 3.0074
1.185 3.0099
1.193 3.0302
1.195 3.0353
1.197 3.0404
1.200 3.048
1.202 3.054

At this point I assumed the data to follow a normal distribution. I determined to make the opening fit the 99th percentile detector. From a z-score table I determined that the 99th percentile is at 2.4 standard deviations above the mean (2.4 std devs above mean is actually 99.2 percentile). The z-score is a dimensionless number obtained by subtracting the population mean from an individual (raw) score and then dividing the difference by the population standard deviation. The z score reveals how many units of the standard deviation a case is above or below the mean. Figure 1 shows a normal distribution with cumulative percentiles and z-scores.



Figure 1. Z-Scores and normal distribution (source: Wikipedia)
This gives the size for the 99th percentile detector of

1.192”+2.4X.008071”=1.211” (3.076 cm)

The number above is the width of a 99th percentile detector. To this number I added 0.005”. This addition was to allow for any inaccuracies in locating the holes for the dowel pins. This addition would allow the dowel pin on each side of a detector to move in by 0.0025”.

The mount was designed with nominal opening for a detector of 1.216” (3.0886 cm).

Unfortunately, designing for 99th percentile detector to ensure that almost all detectors will fit inside the provided opening means that the average detector with a width of 1.192” (3.028 cm) will have 0.024” of side to side travel if the pins are in their correct nominal position.

It turns out that due to the excessive play in the small detectors I will need to redesign the holder.


At this point I created a normal plot to see if the assumption that the data was normally distributed. The normal plot is shown in figure 2.The straight line on the plot indicates normally distributed data. With the limited sample size of the data deviation from a straight line is to be expected. The data is approximately normal. Given the small sample size there is no point to conducting a more formal test for normality.




Figure 2. Normal plot.


Technorati tags:
mechanical design
statistics